Google’s reported Frozen chip would embed parts of Gemini

Google was developing a chip called Frozen v2 that would incorporate elements of Gemini into hardware, Reuters reported, citing The Information. Deployment was under consideration as early as 2028, but the design and amount of model information to embed remained unresolved.

The reported six-to-tenfold efficiency improvement was an ambition, not an independently demonstrated result.

The economic trade-off is how much flexibility can be surrendered for faster or cheaper inference. Specialization pays best when the accelerated workload remains useful long enough to recover development costs. What stays programmable will determine how easily the hardware can accommodate later model changes.